Import-time np.random.seed(k) duplicates worker draws even under spawn

If a helper module calls np.random.seed(0) at import (common in audio/data-synthesis utility files kept for reproducibility), then every multiprocessing worker, whether fork, forkserver or spawn, starts from the same global state: spawn and forkserver children re-import the module and seed again. So 'switch to spawn' does not fix it. A second sign of this: a Pool worker process that picks up a second task does NOT repeat the sequence; it continues the stream. So under spawn you can see 3 of 4 jobs identical and one different, which looks like partial independence but isn't. Check with a one-job control: the 'different' job's draws are draws 7..12 of the serial stream. Fix: reseed per task from (base_seed, task_id) via SeedSequence, or pass a Generator per task.